1 / 20

JMM Special Session

October 14-15, 2011 13th Midwest Optimization Meeting & Workshop on Large Scale Optimization and Applications Hosted by the Fields Institute  http://www.fields.utoronto.ca/programs/scientific/11-12/optimization_mtg/ Friday October 14, 2011, 9:00 a.m. EDT (12.00pm AEST).

Download Presentation

JMM Special Session

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. October 14-15, 2011 13th Midwest Optimization Meeting & Workshop on Large Scale Optimization and Applications Hosted by the Fields Institute  http://www.fields.utoronto.ca/programs/scientific/11-12/optimization_mtg/ Friday October 14, 2011, 9:00 a.m. EDT (12.00pm AEST) JMM Special Session

  2. Midwest Optimization Meeting & Workshop on Large Scale Optimization and Applications October 14th 2011Douglas-Ratchford iterations in the absence of convexity Charles Darwin’s notes Alan Turing’s Enigma Jonathan Borwein FRSC FAA FAASwww.carma.newcastle.edu.au/~jb616 Laureate ProfessorUniversity of Newcastle, NSW Director, Centre for Computer Assisted Research Mathematics and Applications CARMA Revised 16-10-2011: joint work with B. Sims. Thanks also to Ulli Kortenkamp, Fran Aragon, Matt Skerritt and Chris Maitland

  3. THE REST IS SOFTWARE “It was my luck (perhaps my bad luck) to be the world chess champion during the critical years in which computers challenged, then surpassed, human chess players. Before 1994 and after 2004 these duels held little interest.” - Garry Kasparov, 2010 • Likewise much of current Optimization Theory

  4. ABSTRACT • The Douglas-Rachford iteration scheme, introduced half a century ago in connection with nonlinear heat flow problems, aims to find a point common to two or more closed constraint sets. • Convergence is ensured when the sets are convex subsets of a Hilbert space, however, despite the absence of satisfactory theoretical justification, the scheme has been routinely used to successfully solve a diversity of practical optimization or feasibility problems in which one or more of the constraints involved is non-convex. • As a first step toward addressing this deficiency, we provide convergence results for a proto-typical non-convex (phase-recovery) scenario: Finding a point in the intersection of the Euclidean sphere and an affine subspace.

  5. AN INTERACTIVE PRESENTATION • Much of my lecture will be interactive using the interactive geometry package Cinderella and the html applets • www.carma.newcastle.edu.au/~jb616/reflection.html • www.carma.newcastle.edu.au/~jb616/expansion.html • www.carma.newcastle.edu.au/~jb616/lm-june.html

  6. PHASE RECONSTRUCTION Projectors and Reflectors: PA(x) is the metric projection or nearest point and RA(x) reflects in the tangent: x is red x A A PA(x) 2008 Finding exoplanet Fomalhaut in Piscis with projectors RA(x) "All physicists and a good many quite respectable mathematicians are contemptuous about proof." G. H. Hardy (1877-1947) 2007Elser solving Sudoku with reflectors

  7. The story of Hubble’s 1.3mm error in the “upside down” lens (1990) And Kepler’s hunt for exo-planets (launched March 2009) 

  8. The story of Hubble’s 1.3mm error in the “upside down” lens (1990) And Kepler’s hunt for exo-planets (launched March 2009)  A few weeks ago we wrote: “We should add, however, that many Kepler sightings in particular remain to be ‘confirmed.’ Thus one might  legitimately wonder how mathematically robust are the underlying determinations of velocity, imaging, transiting, timing, micro-lensing, etc.? http://experimentalmath.info/blog/2011/09/where-is-everybody/

  9. WHY DOES IT WORK? In a wide variety of large hard problems (protein folding, 3SAT, Sudoku) A is non-convex but DR and “divide and concur” (below) works better than theory can explain. It is: Consider the simplest case of a line B of heighth and the unit circle A. With the iteration becomes For h=0 We prove convergence to one of the two points in A Å B iff we do not start on the vertical axis (where we have chaos). For h>1(infeasible) it is easy to see the iterates go to infinity (vertically). For h=1 we converge to an infeasible point.For h in (0,1) the pictures are lovely but proofs escaped us for 9 months. Two representative Maple pictures follow: An ideal problem for introducing early under-graduates to research, with many many accessible extensions in 2 or 3 dimensions

  10. INTERACTIVE PHASE RECOVERY in CINDERELLA Recall the simplest case of a line B of height h and the unit circle A. With the iteration becomes A Cinderella picture of two steps from (4.2,-0.51) follows:

  11. DIVIDE AND CONCUR Serial (L) and Parallel (R)

  12. CAS+IGP: THE GRIEF IS IN THE GUI Accuracy after taking input from Maple Numerical errors in using double precision

  13. THE ROUTE TO DISCOVERY • Exploration first in Maple and then in Cinderella (SAGE) • - ability to look at orbits/iterations dynamically is great for insight • allows for rapid reinforcement and elaboration of intuition • Decided to look at ODE analogues • - and their linearizations • hoped for Lyapunov like results • Searched literature for a discrete version • - found Perron’s work

  14. THE BASIS OF THE PROOF Explains spin for height in (0,1)

  15. WHAT WE CAN NOW SHOW

  16. ALGORITHM APPEARS TO BE STABLE

  17. THREE AND HIGHER DIMENSIONS

  18. AN “EVEN SIMPLER” CASE ≈ 1.51 RANDOM POINTS Intersection at

  19. COMMENTS and OPEN QUESTIONS • As noted, the method parallelizes very well. • Work out rates in convex case? • Why does alternating projection (no reflection) work well • for optical aberration but not phase reconstruction? • Show rigorously global convergence • -in the appropriate basins? • Extend analysis to more general pairs of sets (and CAT (0) metrics) • -even the half-line case is much more complex • - as I may now demo.

  20. REFERENCES • [1] Jonathan M. Borwein and Brailey Sims, “The Douglas-Rachford algorithm in the absence of convexity.” Chapter 6, pp. 93–109 in Fixed-Point Algorithms for Inverse Problems in Science and Engineering in Springer Optimization and Its Applications, vol. 49, 2011. • http://www.carma.newcastle.edu.au/~jb616/dr.pdf • [2] J. M. Borwein and J. Vanderwerff, Convex Functions: Constructions, Characterizations and Counterexamples. Encyclopedia of Mathematics and Applications, 109 Cambridge Univ Press, 2010. • http://projects.cs.dal.ca/ddrive/ConvexFunctions/ • [3] V. Elser, I. Rankenburg, and P. Thibault, “Searching with iterated maps,” Proc. National Academy of Sciences, 104 (2007), 418–423. • [4] J.M. Borwein and R.L. Luke, ``Duality and Convex • Programming," pp. 229–270 in Handbook of Mathematical Methods • in Imaging, O. Scherzer (Editor-in-Chief), Springer-Verlag. E-pub. 2010.

More Related